ContextCore for AI editors

Official ContextCore plugin for Cursor and Claude Code — product context over MCP, the ContextCore playbook skill, and the init command. Claude Desktop uses Custom Connectors or an optional .mcpb Extension.

Editor matrix: EDITORS.md — Cursor, Claude Code plugin, Claude Desktop Connectors, .mcpb, stdio, CLI.

Give agents secure access to your ContextCore project context — vision, audience, scope, tech decisions, rules, skills, risks, and open questions — plus write tools to propose updates.

With a read-scoped token or OAuth grant, the agent reads published context only. With write access, it can also propose context, rules, skills, risks, and problems (landing in New for review unless you have direct-publish permission).

The agent gets exactly the context you can see. Access is enforced server-side by ContextCore's org/team/visibility federation.

Quick start

Editor	Action
Cursor	Customize → search ContextCore → Install
Claude Code	`claude plugin marketplace add babaVC/contextcore-mcp` → `claude plugin install contextcore@contextcore-mcp`
Claude Desktop	Customize → Connectors → `https://cloud.contextcore.md/mcp`
Details	EDITORS.md

Cursor + Claude Code plugins (this repo)

This repo is the ContextCore editor plugin for Cursor and Claude Code — shared skills, commands, and MCP wiring. Claude Desktop uses Custom Connectors (same gateway URL) or the optional .mcpb in desktop-extension/. See EDITORS.md.

Cursor Directory: cursor.directory/plugins/contextcore (live). Official marketplace: submitted 2026-09-03 — awaiting review. Claude Code directory: submitted 2026-09-08 — pending review. See SUBMISSION.md.

Local Cursor test: clone to ~/.cursor/plugins/local/contextcore and reload Cursor.

Claude Code test:

claude plugin marketplace add babaVC/contextcore-mcp
claude plugin install contextcore@contextcore-mcp

Claude Code (plugin or MCP-only)

Plugin (recommended): see above — bundled skill + /contextcore:contextcore-init.

MCP only:

claude mcp add --transport http contextcore https://cloud.contextcore.md/mcp

Browser OAuth on the first MCP call. Paste Set up ContextCore and initialize it. — same Initialize procedure as /contextcore-init.

More: EDITORS.md · MCP user guide

MCP servers in the plugin (Cursor)

Primary — contextcore (remote / OAuth). No token to paste. The first request opens browser sign-in and consent. Configure CONTEXTCORE_MCP_URL (default https://cloud.contextcore.md/mcp).

Fallback — contextcore-stdio. For headless setups or when you prefer a Personal Access Token. Set CONTEXTCORE_TOKEN and optionally CONTEXTCORE_API_URL. Uses this repo's npm package via npx.

If you only use OAuth, leave CONTEXTCORE_TOKEN unset and disable the contextcore-stdio server in Cursor MCP settings — otherwise it shows as failed until a token is configured.

On-prem: point CONTEXTCORE_MCP_URL and CONTEXTCORE_API_URL at your instance.

What your agent gets

Read tools (default):

Tool	What it returns
`list_projects`	Projects you can access (id, name, tagline)
`get_manifest`	Resolved map, gaps, counts, watch roll-ups
`get_context_core`	A project's full context as one Markdown document
`get_context_tree`	File layout of the Core export
`read_component`	One component (e.g. `vision`, `scope`, `tech`) as Markdown
`read_rule` / `read_skill`	One rule or skill by slug — use `contextcore` for the platform playbook
`list_components`	Component list for a project
`search_context`	Keyword matches across pieces, rules, skills, risks, and problems

Write tools (write-scoped token or OAuth grant with propose changes):

Tool	What it does
`propose_context`	Propose a context piece or update (with `why` evidence)
`propose_rule` / `propose_skill`	Propose agent guidance assets
`flag_risk` / `flag_problem`	Propose watch items
`list_suggestions`	List pending items in New
`review`	Approve or reject a suggestion

Tools are fetched from the server at startup — no package upgrade needed when the gateway adds tools.

Standalone stdio package (non-Cursor editors)

The contextcore-mcp npm package (bin/cli.js in this repo) is a JSON-RPC proxy for stdio-only MCP clients. The Cursor plugin reuses it as contextcore-stdio; other editors can use it directly:

{
  "mcpServers": {
    "contextcore": {
      "command": "npx",
      "args": ["-y", "contextcore-mcp"],
      "env": { "CONTEXTCORE_TOKEN": "YOUR_TOKEN" }
    }
  }
}

For sync/review/propose from the terminal, see contextcore-cli.

Manual MCP config (no plugin)

Remote OAuth — works in Cursor, Claude Code, and other HTTP MCP clients:

{
  "mcpServers": {
    "contextcore": {
      "url": "https://cloud.contextcore.md/mcp"
    }
  }
}

Mint a token under Account → Integrations only if you need stdio or automation.

Sync workflow (maintainers)

Playbook copy lives in the main ContextCore app repo, not here. skills/ and commands/ are generated — do not hand-edit.

Edit src/constants/mcpPlaybook.js in context-os (private app repo)
Run npm run sync:cursor-plugin from that repo

That writes skills/contextcore/SKILL.md, commands/contextcore-init.md, and dogfoods the command into context-os/.cursor/commands/.

Configuration

Variable / env	Required	Default	Description
`CONTEXTCORE_MCP_URL`	no	`https://cloud.contextcore.md/mcp`	Remote MCP gateway (plugin variable)
`CONTEXTCORE_API_URL`	no	`https://cloud.contextcore.md`	App origin for stdio proxy
`CONTEXTCORE_TOKEN`	stdio only	—	Personal Access Token from ContextCore

Legacy CONTEXTOS_TOKEN / CONTEXTOS_API_URL still work in the stdio client.

Security

OAuth grant or token maps to one ContextCore user. Federation gates match the web app.
Write proposals land in New unless you have direct-publish permission.
Token hashes only — revoke anytime from Account → Integrations.
Never commit tokens. Use plugin variables or ${env:…} in config.

See contextcore.md/trust.

License

MIT

ContextCore for AI editors